Video transport over ad hoc networks: multistream coding ...catt.poly.edu/~panwar/publications/video...

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 21, NO. 10, DECEMBER 2003 1721

Video Transport Over Ad Hoc Networks: MultistreamCoding With Multipath Transport

Shiwen Mao, Student Member, IEEE, Shunan Lin, Student Member, IEEE, Shivendra S. Panwar, Senior Member, IEEE,Yao Wang, Senior Member, IEEE, and Emre Celebi, Student Member, IEEE

Abstract—Enabling video transport over ad hoc networks ismore challenging than over other wireless networks. The wirelesslinks in an ad hoc network are highly error prone and can go downfrequently because of node mobility, interference, channel fading,and the lack of infrastructure. However, the mesh topology of adhoc networks implies that it is possible to establish multiple pathsbetween a source and a destination. Indeed, multipath transportprovides an extra degree of freedom in designing error resilientvideo coding and transport schemes.

In this paper, we propose to combine multistream coding withmultipath transport, to show that, in addition to traditional errorcontrol techniques, path diversity provides an effective means tocombat transmission error in ad hoc networks. The schemes thatwe have examined are: 1) feedback based reference picture selec-tion; 2) layered coding with selective automatic repeat request;and 3) multiple description motion compensation coding. All thesetechniques are based on the motion compensated prediction tech-nique found in modern video coding standards. We studied the per-formance of these three schemes via extensive simulations usingboth Markov channel models and OPNET Modeler. To further val-idate the viability and performance advantages of these schemes,we implemented an ad hoc multiple path video streaming testbedusing notebook computers and IEEE 802.11b cards. The resultsshow that great improvement in video quality can be achieved overthe standard schemes with limited additional cost. Each of thesethree video coding/transport techniques is best suited for a par-ticular environment, depending on the availability of a feedbackchannel, the end-to-end delay constraint, and the error character-istics of the paths.

Index Terms—Ad hoc networks, error resilience, IEEE 802.11,multipath transport, video transport, wireless networks.

I. INTRODUCTION

A D HOC NETWORKS are multihop wireless networkswithout a preinstalled infrastructure. They can be de-

ployed instantly in situations where infrastructure is unavailable(e.g., disaster recovery), or where infrastructure is difficult to

Manuscript received October 15, 2002; revised May 1, 2003. This workwas supported in part by the National Science Foundation under Grant ANI0081375, in part by the New York State Center for Advanced Technology inTelecommunications (CATT) and the Wireless Internet Center for AdvancedTechnology (WICAT), Polytechnic University, Brooklyn, NY. This paper waspresented in part at the IEEE International Conference on Multimedia andExpo (ICME), Tokyo, Japan, August 2001, and at the Vehicular TechnologyConference, Atlantic City, NJ, October 2001.

S. Mao, S. Lin, S. S. Panwar, and Y. Wang are with the Department of Elec-trical and Computer Engineering, Polytechnic University, Brooklyn, NY 11373USA (e-mail: [email protected]; [email protected]; [email protected];[email protected]).

E. Celebi is with the Department of Computer and Information Science, Poly-technic University, Brooklyn, NY, 11373 USA (e-mail: [email protected].).

Digital Object Identifier 10.1109/JSAC.2003.815965

install (e.g., battlefields). It is maturing as a means to provideubiquitous untethered communication. With the increase bothin the bandwidth of wireless channels and in the computingpower of mobile devices, it is expected that video service willbe offered over ad hoc networks in the near future.

Ad hoc networks pose a great challenge to video transport.There is no fixed infrastructure and the topology is frequentlychanging due to node mobility. Therefore, links are continu-ously established and broken. The availability and quality of alink further fluctuates due to channel fading and interferencefrom other transmitting users. In addition, an end-to-end pathconsists of a number of wireless links. Thus, transmission lossin ad hoc networks is more frequent than that in wireless net-works with single-hop wireless paths connecting nodes to thewireline infrastructure. The most popular media access control(MAC) scheme, the carrier sensing multiple access/collisionavoidance (CSMA/CA) scheme [1], is designed for best-effortdata. It provides no hard guarantees for a session’s band-width and delay. Although bandwidth reservation is possiblewith MAC schemes based on time-division multiple-access(TDMA) or code-division multiple-access (CDMA), practicalimplementations of these schemes are nontrivial because ofthe synchronization or code assignment problems when nodemobility is allowed [2].

Video transport typically requires stringent bandwidth anddelay guarantees. However, it is very hard to maintain anend-to-end route, which is both stable and has enough band-width in an ad hoc network. Furthermore, compressed videois susceptible to transmission errors. For example, a singlebit error often causes a loss of synchronization when variablelength coding (VLC) is used. Moreover, the motion compen-sated prediction (MCP) technique is widely used in modernvideo coding standards. In MCP, a frame is first predicted froma previous coded frame (called reference picture) and then theprediction error is encoded and transmitted. Although MCPachieves high coding efficiency by exploiting the temporalcorrelation between adjacent frames, it makes the reconstruc-tion of a frame depend on the successful reconstruction ofits reference picture. Without effective error protection andconcealment, a lost packet in a frame can cause not only errorwithin this frame, but also errors in many following frames,even when all the following frames are correctly received [3].

Given the error-prone nature of ad hoc network paths andthe susceptibility of compressed video to transmission errors,effective error control is needed. Traditional techniques, in-cluding forward error correction (FEC) and automatic repeat

0733-8716/03$17.00 © 2003 IEEE

1722 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 21, NO. 10, DECEMBER 2003

request (ARQ), must be adapted to take into considerationthe delay constraint and the error propagation problem [4].In ad hoc networks, wireless links break down the traditionalconcept of topology, which is not constrained by physical cableconnections anymore. Although user mobility makes linksvolatile, it provides variability of topology. On the one hand,a link may break when nodes move away from each other. Onthe other hand, it is possible to quickly find new routes formedin a new topology. Furthermore, the mesh topology of ad hocnetworks implies the existence of multiple routes between twonodes. Given multiple paths, a video stream can be dividedinto multiple substreams and each substream is sent on oneof the paths. If these paths are disjoint, the losses experiencedby the substreams would be relatively independent. Therefore,better error resilience can be achieved when traffic dispersion isperformed appropriately and with effective error control for thesubstreams. In a manner similar to multiantenna diversity thatimproves the capacity of wireless networks, path diversity canalso be exploited to improve the capacity of ad hoc networks.Indeed, multipath transport (MPT) provides an extra degree offreedom in designing video coding and transport schemes.

In this paper, we propose three MCP-based video transporttechniques for mobile ad hoc networks. These schemes take ad-vantage of path diversity to achieve better performance. Com-pared with the coding methods considered in our previous work[5], these techniques can achieve significantly higher coding ef-ficiency and are compliant with the and MPEG seriesstandards (possibly with simple modifications). The techniquesthat we have examined include:

1) a feedback based reference picture selection scheme(RPS) [6];

2) a layered coding (LC) with selective ARQ scheme (LCwith ARQ) [7];

3) a multiple description motion compensation codingscheme (MDMC) [8].

We studied the performance of these three schemes via atop-downapproach. First ,we used a popular Markov link model[6], [9], where lower layer detail is embodied in the bursty errorsgenerated. This simple model enables us to examine the systemperformance over a wide range of pack loss rates and loss pat-terns. Next, lower layer details, including user mobility, multi-path routing, and the MAC layer are taken into account in theOPNET simulations [10], which provide a more realistic viewof the impact of these factors on the system performance. Fur-thermore, we implemented an ad hoc video streaming testbedusing notebook computers with IEEE 802.11b cards. This fur-ther validates the viability and performance advantages of theseschemes. The results of our experiments show that video trans-port is viable in ad hoc networks given careful cross-layer de-sign. Combining multistream coding with MPT improves videoquality, as compared with traditional schemes where a singlepath is used. Each of these three techniques is best suited for aparticular environment, depending on the availability of feed-back channels, the end-to-end delay constraint, and the errorcharacteristics of the paths.

The remainder of the paper is organized as follows. InSection II, we present the general architecture of multistream

Fig. 1. General architecture of the proposed system using multistream codingand multipath transport.

video coding with MPT. We also discuss related issues andprior work in this section. Next, the three multistream codingand MPT schemes are discussed in Section III. Sections IV andV present the performance study of these schemes using theMarkov models and OPNET Modeler, respectively. Our exper-imental results with an ad hoc network video streaming testbedare reported in Section VI. Section VII provides discussion andconclusions.

II. V IDEO TRANSPORTWITH PATH DIVERSITY

A. General Architecture

The general architecture of the proposed system is shownin Fig. 1. In the architecture, a multipath routing layer sets up

paths between the source and destination, each with a setof quality-of-service (QoS) parameters in terms of bandwidth,delay, and loss probabilities. The transport layer continuouslymonitors path QoS parameters and returns such information tothe sender. Based on path quality information, the encoder gen-erates substreams. Packets from the substreams are dispersedby the traffic allocator among the paths. At the receiver,packets arriving from all the paths are put into a resequencingbuffer where they are reassembled into substreams after apreset playout delay. Some or all the packets assigned to a pathmay be lost or overdue. Limited retransmission of lost packetsmay or may not be invoked, depending on the encoding schemeand the end-to-end delay constraint. The decoder will attemptto reconstruct a video sequence from the received substreams.

B. Related Issues and Challenges

A key to the success of the proposed system is the close inter-action between the source coder and the transport layer, whichentails careful cross-layer design. We will highlight this inter-action in the following discussion.

1) Multistream Video Coding:For MPT to be helpful forsending compressed video, one must carefully design thevideo coder to generate substreams so that the loss in onesubstream does not adversely affect the decoding of othersubstreams. However, this relative independence between thesubstreams should not be obtained at a great expense of codingefficiency. Therefore, the multistream encoder should striveto achieve a good tradeoff between the coding efficiency anderror resilience. In addition, one must consider what is feasiblein terms of transport layer error control, when designing thesource coder.

MAO et al.: VIDEO TRANSPORT OVER Ad Hoc NETWORKS: MULTISTREAM CODING WITH MULTIPATH TRANSPORT 1723

Obviously, one way to generate multiple substreams is touse a standard video codec and split the resulting bitstreaminto multiple substreams. An intelligent splitting scheme isneeded to split the bit stream at the boundary of independentlydecodable units. Otherwise, a lost substream will make thereceived ones from other paths useless. A simple way toaccomplish this is to send the frames to the paths in a roundrobin manner, e.g., all odd frames are sent to path 1 andall even frames are sent to path 2. In order to completelyavoid the dependency between substreams, the frames senton one path should be predictively coded with respect to theframes on the same path only. This method is in fact an optionavailable in the standard [video redundancy coding(VRC)] [11]. However, compared with predicting a frame fromits immediate neighbor, VRC requires significantly higher bitrates. Also, although this method can prevent the loss in onepath from affecting frames in the other path, error propagationstill exists within frames in the same path. In this paper, weintroduce a feedback based reference picture selection method,which circumvents these two problems of VRC. This methodis introduced in Section III-A.

Another natural way of generating multiple streams is byusinglayered video coding, which is very useful in coping withthe heterogeneity of user access rates, in network link capaci-ties, and in link reliability. A layered coder encodes video intoseveral layers. The base layer (BL), which includes the cru-cial part of the video frames, guarantees a basic display quality.Each enhancement layer (EL) correctly received improves thevideo quality. But without the BL, video frames cannot be re-constructed sufficiently. Usually, EL packets may be droppedat a congested node to protect BL packets, and BL packets arebetter protected with FEC or ARQ [12]. When combined withMPT, it is desirable to transmit the BL substream on the bestroute. The source may sort the paths according to their loss char-acteristics, inferred from QoS feedback (e.g., receiver reports inthe real-time transport protocol (RTP) [13]). Alternatively, themultipath routing layer may organize the route cache accordingto some performance metrics (number of hops, mean loss ratein the last time window, etc.). In this paper, we consider an ap-proach that protects the BL by retransmitting lost BL packetson the path carrying the EL packets. This method is describedin Section III-B.

Instead of generating substreams that are unequal in theirimportance, multiple description coding(MDC) generatesmultiple equally important streams, each giving a low butacceptable quality. A high-quality reconstruction is decodablefrom all bit streams together, while a lower, but still acceptablequality reconstruction is achievable if only one stream is re-ceived. The correlation among the substreams introduced at theencoder makes it possible to partially recover lost informationof one substream, using information carried in other correctlyreceived substreams. However, such a correlation limits theachievable coding efficiency, as compared with a conventionalcoder designed to maximize it. An excellent review of thetheoretical bounds and proposed MDC algorithms can be foundin [14]. In designing a MCP-based multiple description (MD)video codec, a key challenge is how to control the mismatch

between the reference frames used in the encoder and thoseused in the decoder caused by transmission errors. Amongseveral MD video coding schemes proposed so far [8], [11],[15], [16], we chose the MDMC method (Section III-C),because it outperformed other MD video coding methodsin our previous studies [8]. With MDC, the transport layerdesign can be simpler than with layered coding. Because all thedescriptions are equally important, the transport layer does notneed to protect one stream more than another. Also, becauseeach description alone can provide a low but acceptable quality,no retransmission is required, making MDC more suitablefor applications with stringent delay requirements. Thoughthe focus of this paper is on point-to-point communications,MDC can also be considered for multicast applications, whereretransmissions are best avoided [17].

2) Multipath Transport: MPT has been studied in the past inwireline networks for: i) increased aggregate capacity; ii) betterload balancing; and iii) path redundancy for failure recovery[18]–[20]. The research effort on MPT can be roughly dividedinto the following two categories.

a) Multipath routing, which focuses on finding multipleroutes for a source-destination pair, and on how to selecta maximally disjoint set of routes from the multipleroutes found [21]–[24];

b) Traffic dispersion, which focuses on how to allocate trafficto multiple end-to-end routes [25], [26]. Generally, trafficdispersion can be performed with different granularities.[27] is an excellent survey of this topic.

The particular communication environment of wireless adhoc networks makes MPT very appealing. In ad hoc networks:i) individual links may not have adequate capacity to supporta high bandwidth service1 ; ii) a high loss rate is typical; andiii) links are unreliable. MPT can provide larger aggregate band-width and load balancing for ad hoc video applications. In ad-dition, the path diversity inherent in MPT can provide bettererror resilience performance. Furthermore, many of the ad hocnetwork routing protocols, e.g., dynamic source routing (DSR)[28], ad hoc on-demand distance vector (AODV) [29], and zonerouting protocol (ZRP) [30], are able to return multiple paths inresponse to a route query. Multipath routing can be implementedby extending these protocols with limited additional complexity.

There are many challenges in supporting MPT in ad hoc net-works. First, from multiple paths returned by a route query, therouting process should select a set of maximally disjoint paths.Shared or nearby links of the paths could make the loss pro-cesses of the substreams correlated, which reduces the benefitof using MPT [31]. Algorithms for finding disjoint paths are pre-sented in [21], [22]. Second, finding and maintaining multiplepaths requires higher complexity and may cause additional over-head on traffic load (e.g., more route replies received). However,caching multiple routes to any destination allows prompt reac-tion to route changes. If a backup path is found in the cache,

1Although in some cases the nominal bandwidth of a wireless link is com-parable to that of a wireline link, the available bandwidth may vary with signalstrength as in IEEE 802.11b. In addition, capacity lost due to protocol overheadin ad hoc networks is much higher than that in wireline networks (e.g., RTS,CTS, ACK packets, and the 30-byte frame header in IEEE 802.11b).


there is no need to send new route queries. Rerouting delay androuting overhead may be reduced in this case. These problemsshould be addressed carefully in the design of a multipath trans-port protocol to balance its benefits. Third, a problem inherentin MPT is the additional delay and complexity in packet re-sequencing. Previous work shows that resequencing delay andbuffer requirement are moderate if the traffic allocator in Fig. 1is carefully designed [32], [33].

C. Related Work

Due to the availability of a variety of network accesstechnologies, as well as the reduction in their costs, there isstrong interest in taking advantage of multihomed hosts to getincreased aggregate bandwidth and higher reliability. Proposalsin the transport layer include [20], [34]–[36]. In [34], a protocolcalled meta-transmission control protocol (meta-TCP) main-taining multiple TCP connections for a session was designedfor data transport. The stream control transmission protocol(SCTP) [20] was initially designed for reliable delivery ofsignaling messages in Internet protocol (IP) networks usingpath redundancy. There are now proposals to adapt it for datatraffic in the Internet and in wireless networks [35], [36].These papers focus on the higher aggregate usable bandwidthobtained and on how to perform TCP congestion control overmultiple paths. Multiflow management can also be carried outat the application layer. In [5] and [37], an extension of the RTP[13], called meta-RTP, was proposed. Meta-RTP sits on top ofRTP in the protocol stack, performing traffic allocation at thesender and resequencing at the receiver for real-time sessions.

Recently, several interesting proposals on delivering audioand video over Internet and wireless networks using multiplepaths have been introduced. The study in [5] and [37] was,to the best of our knowledge, the first to investigate imageand video transport using MPT in a multihop wireless radionetwork. Although it provided some very useful insights,the coders considered there treated individual frames of avideo sequence independently, and consequently are not veryefficient. There are several interesting papers on applying MPTfor Internet multimedia streaming. In [38], MDC is combinedwith path diversity for video streaming in the Internet. Afour-state model is proposed to capture the distortion behaviorof a MD source. The problem of MD video downloading incontent distribution networks (CDN) using a number of serversis studied in [39]. It is reported that 20% to 40% reductions indistortion can be achieved by using thismany-to-oneapproach.Similarly, it is shown in [40] that using multiple senders andFEC in data downloading effectively reduces packet lossrates. An interesting study on realtime multistream voicecommunication through disjoint paths is given in [41], wheremultiple redundant descriptions of a voice stream are sent overpaths provided by different Internet service providers. Bothsignificant reductions in end-to-end latency and packet lossrate are observed. In recent work [42], the RPS scheme in [6]was extended by using rate-distortion optimized long memoryreference picture selection, and using probes for path statusprediction.

This paper differs from previous work discussed thus far inmany aspects. First, we focus on video transport, as comparedwith general elastic data transport using TCP [20], [34]–[36].We perform traffic partitioning in the application layer anduse user datagram protocol (UDP) in the transport layer. Weperform traffic dispersion on the substream level for the resultsshown in this paper.2 Compared with meta-RTP [5], [37] ourtransport schemes require no integrity in each substream andare more flexible. Second, we study multipath video transportin ad hoc networks, while prior work focuses on Internet videostreaming [38]–[42]. It is much more challenging to transportvideo in ad hoc networks than in a wireline network, e.g.,the Internet, as discussed in Section I. Moreover, there is thewell-known assumption that all packet losses in the Internetare caused by congestion [43], and MPT is mainly used in theInternet to alleviate congestion (i.e., load balancing). In an adhoc network, in addition to congestion in mobile nodes, packetsare also lost because the wireless links are unreliable. Therefore,the benefit of using MPT, in addition to load balancing, iserror resilience through path diversity. Since the up and downstatus of the paths are relatively independent of each other,it is possible to apply efficient error control exploiting thisfeature to improve video quality. Third, the multistream videocoding schemes we proposed are all MCP-based with highcoding efficiency and are more compliant with modern videocoding standards, as compared with [5], [37]. MDMC is a newmultiple description video coding technique and [16] describesits algorithm and its performance using abstract channel models.In this paper, we study the performance of MDMC in adhoc networks and performed extensive performance studiesof MDMC under a more realistic network setting. Fourth,we extend the DSR protocol [28] to support multiple pathrouting. With our extension, multiple maximally disjoint routesare selected from all the routes returned by a route query,with only limited increase in the routing overhead. Fifth, forperformance evaluation, we adopt a realistic model whichincludes all the layers except the physical layer using OPNETModeler [10]. We believe this cross-layer model provides arealistic view for video transport over ad hoc networks.

In terms of implementation work, a number of ad hoc testbedshave been built recently [44], [45]. These mainly focus on theperformance of ad hoc routing protocols, physical layer charac-teristics, scalability issues, and integration of ad hoc networkswith the Internet for data transport. In [46], a firewall is insertedbetween the source and destination, which drops video packetsaccording to a Markov channel model [9]. So far as we know,the testbed we developed is the first effort in combining mul-tistream video coding and MPT for video transport in ad hocnetworks.

III. PROPOSEDVIDEO CODING AND TRANSPORTSCHEMES

One of the challenges when utilizing path diversity for videotransmission is how to generate multiple coded substreams

2Finer packet-level traffic dispersion schemes can be supported by ourschemes when more than two paths are available.


to feed the multiple paths. We consider three types of codingschemes that differ in terms of their requirements for thetransport-layer support. These three schemes are all built ontop of the block-based hybrid coding framework using MCPand discrete cosine transform (DCT), which is employed by allexisting video coding standards. This way, the loss of codingefficiency is limited and the source codec can be implementedby introducing minimal modifications to codecs followingexisting standards. We present these three methods separatelyin the subsequent subsections, followed by comparison anddiscussion.

A. Feedback Based Reference Picture Selection

As discussed above, one of the main challenges in MCP-based video coding for ad hoc networks is how to limit theextent of error propagation caused by loss on a bad path, andyet minimize the loss in coding efficiency. As mentioned inSection II-B, one simple approach to generate two substreams isthe VRC option in the standard [11], which codes theeven and odd frames as two separate substreams and performtemporal prediction within each substream. However, comparedwith predicting a frame from its immediate neighbor, VRC re-quires significantly higher bit rates. Also, although this methodcan prevent the loss in one path from affecting frames in theother path, error propagation still exists within the same path.We note that there is no reason to forbid one path from usinganother path’s frames as reference if all the paths are good. Mo-tivated by this observation, we propose to choose the referenceframes as follows: based on feedback and predicted path status,always choose the last frame that is believed to have been cor-rectly received as the reference frame.

Specifically, we sent the coded frames on separate paths. Themapping of frames to paths depends on the available bandwidthon each path. For example, in the two-path case, if both pathshave the same bandwidth, then even frames are sent on path 1,and odd frames on path 2. We assume that a feedback messageis sent for each frame by the decoder. If any packet in a frameis lost, the decoder sends a negative feedback (NACK) for thatframe. Otherwise, it sends a positive feedback (ACK). The feed-back information for a frame may be sent on the same path asthe frame, or on a different path. An encoder receives the feed-back message for frame when it is coding frame ,where round-trip time (RTT) is measured in frame intervals.

Furthermore, once a NACK is received for a frame deliveredon one path, we assume that the path remains “bad” until anACK is received. Similarly, we assume the path stays in the“good” status until a NACK is received. When encoding a newframe, the encoder deduces the last correctly decoded frame,based on the feedback messages received up to this time, anduses that frame as the reference frame. This scheme works wellwhen the loss of a path is bursty, which is typical in ad hocnetworks. A more sophisticated scheme may adopt a thresholdfor the NACKs and ACKs for a given window of time and switchthe reference frame only when the threshold is exceeded.

Fig. 2 is an example of the proposed RPS scheme where. When NACK(1) is received at the time when frame 4

is being encoded, the encoder knows that frames 2 and 3 cannot

Fig. 2. Illustration of the RPS scheme. The arrow associated with each frameindicates the reference used in coding that frame.

be decoded correctly due to error propagation. Therefore, frame0 is chosen as the reference for frame 4 and path 2 is set to “bad”status. When frame 6 is coded, the encoder uses frame 4 insteadof frame 5 as reference frame, because path 2 is still in the “bad”status. When ACK(7) is received, path 2 is changed to “good”status. Frame 9 is then chosen as the reference of frame 10.

The RPS scheme offers a good tradeoff between codingefficiency and error resilience. When both paths are good, RPSuses the immediate neighboring frame as reference, therebyachieving the highest possible prediction gain and codingefficiency. When one path is bad, the encoder avoids using anyframes that are affected by path errors, thereby minimizing theerror propagation period. RPS has higher coding efficiencythan the schemes using fixed prediction distances [11] anderror propagation in each substream is effectively suppressed.These improvements are achieved by using a frame buffer tostore several previous coded frames as possible references,and by using feedback for path status prediction. Note that inthis scheme feedback is used to control the operation of theencoder. No retransmission is invoked.

B. Layered Coding With Selective ARQ

This is a scheme using layered video coding. With thisscheme, a raw video stream is coded into two layers, a BLand an EL. We follow the SNR profile in the H.263 standard[48] when generating the layers. A BL frame is encoded usingthe standard predictive video coding technique. Note thatbecause the BL coding uses only the previous BL picture forprediction, this coding method has a lower coding efficiencythan a standard single layer coder. This loss in coding efficiencyis, however, justified by increased error resilience: a lost ELpacket will not affect the BL pictures. Good quality is, thus,guaranteed if the BL packets are delivered error-free or at a verylow loss rate. There are three prediction options in the H.263standard for enhancement layer coding: UPWARD prediction,in which the BL reconstruction of current frame is used as theprediction of the enhancement layer, FORWARD prediction, inwhich the enhancement layer reconstruction from the previousframe is used as the prediction, and BIDIRECTION prediction,in which the average of BL reconstruction and enhancementlayer reconstruction is used. The LC with ARQ codec selectsfrom the three prediction options the one that has the bestcoding gain. Although this approach is optimal in terms ofcoding efficiency for the enhancement layer, error propagationcan still occur in the EL pictures.


Fig. 3. Two-path layered video transmission model with end-to-end ARQ forBL packets.

Given two paths, the traffic allocator sends the BL packetson one path (the better path in terms of loss probability whenthe two paths are asymmetric) and the EL packets on the otherpath. The receiver sends selective ARQ requests to the sender toreport BL packet losses. To increase the reliability of the feed-back, a copy of the ARQ request is sent on both paths. Whenthe sender receives an ARQ request, it retransmits the requestedBL packet on the EL path, as illustrated in Fig. 3. The transmis-sion bit rate for the EL may vary with the bit rate spent on BLretransmission. For video streaming applications where videois typically pre-encoded off-line, a simple rate control methodis used for the EL path: when a BL packet is retransmitted onthe EL path, one or more EL packets are dropped to satisfy thetarget transmission rate on the EL path.

Our observations show that a multiple hop wireless path be-haves more in an on–off fashion with bursty packet losses. Ifthere is a BL packet loss, the BL route is most likely to bebroken. Moreover, if the loss is caused by congestion at an inter-mediate node, using the BL path for retransmission may makethe congestion more severe. If disjoint paths are used, path di-versity implies that when the BL path is down or congested, itis less likely that the EL path is also down or congested. There-fore, retransmission using the EL path is likely to have a highersuccess probability and lower delay.

As discussed in Section II-B, we assume either the sendercontinuously estimates the states of the paths based on receivedARQ requests or QoS reports, or the multipath routing processorders the paths according to their loss characteristics. In thefirst case, a burst of ARQ requests received at the sender im-plies that the BL path is in a “bad” state. If the inferred EL pathstate is better, the sender may switch the paths. In the latter case,an intermediate node may send an error report back to the sourceafter it drops a packet (e.g., an error report in DSR [28], or an In-ternet control message protocol (ICMP) unreachable error mes-sage [47]). These error reports will trigger the routing processto reorder the paths or initiate a new route query.

C. Multiple Description Motion Compensation

Unlike the above two techniques, MDMC is a multipledescription coding scheme which does not depend on theavailability of feedback channels. Because paths in ad hocnetworks change between “up” and “down” state very often,each description experiences bursty packet losses. Therefore,

TABLE ICOMPARISON OF THETHREE SCHEMES

we employ thepacket-loss modeof MDMC presented in [10].It uses a linear superposition of two predictions from twopreviously coded frames. In the MDMC encoder, the centralprediction is obtained by

(1)

where and are motion compensated pre-dicted signals constructed from two previously encoded frames

and , respectively. The central predictionerror is quantized by quantizerto . The quantized prediction error and motion vectors foreven frames are sent on one path, and those for odd frames aresent on another path. In the decoder, if frame is received,frame is reconstructed using

(2)

where represents motion compensated prediction fromdecoded frame .

If frame is damaged but frame is received, thedecoder only uses the reconstructed frame for prediction.To circumvent the mismatch between the predicted frames usedin the encoder and the decoder, the signal

is quantized byanother quantizer , which is typically coarser than ,and the output is sent along with other information onframe . Now, when frame is damaged, the side decoderreconstructs frame using

(3)

In addition, the lost frame is estimated using

(4)The MDMC codec offers a tradeoff between redundancy and

distortion over a wide range by varying the coder parameters(the predictor coefficient and the quantization parameterof ). The efficiency of a MDMC codec depends on theselection of the parameters, which in turn depends on theestimation of the channel’s error characteristics. There isonly one additional buffer needed in MDMC compared withconventional codecs that use only one previous frame forprediction.

D. Comparison and Discussion

The three schemes have their respective advantages anddisadvantages. Depending on the availability of a feedbackchannel, the delay constraint, and the error characteristics ofthe established paths, one technique may be better suited for anapplication than another. A comparison of these three schemesis given in Table I.


RPS is applicable when feedback channels are available. Theredundancy depends on the distance between a current frameand its reference frame, which in turn depends on the packetloss rate and the RTT. When the paths are error-free, RPS hasthe highest encoding efficiency. Compared with ARQ-basedschemes, there is no decoding delay incurred but additionalbuffers are still needed.

LC with ARQ is suitable when feedback channels are avail-able and the application is such that the latency caused by re-transmission is tolerable. The redundancy of this scheme comesfrom the fact that a frame is predicted from the BL reconstruc-tion of the reference frame. It is difficult to control the amountof the redundancy introduced, which is more than the amountof redundancy introduced in the MDMC coder, when operatingunder the chosen set of parameters. This is why the LC approachhas the lowest quality when packet loss rate is low. However,when the packet loss rate is high, this method can usually deliverthe BL successfully, thus providing better video quality than theother two proposed schemes, at the cost of extra delay. The ad-ditional delay is at least RTT.

MDMC, unlike the other two, does not need feedback, nordoes it incur additional decoding delay. It is easier to controlthe redundancy in MDMC by changing the predictors and theside quantizer. The redundancy can be achieved in a wider rangethan the above two schemes (even though the parameters of theMDMC coder are fixed in all the simulation studies reportedhere). Since MDMC needs no feedback, the video can be pre-en-coded, which is desirable for video streaming applications. Notethat this is not possible with the RPS scheme. The challengewith MDMC is how to adapt the coding parameters based onthe error characteristics of the paths so that the added redun-dancy is appropriate.

IV. PERFORMANCESTUDY USING MARKOV MODELS

In this section, we report on performance studies of the pro-posed schemes. The challenge is that the problem requires cross-layer treatment with a large set of parameters. To simplify theproblem and focus on the key issues, we first study the perfor-mance of the schemes using Markov link models. The lowerlayer details are hidden and their impact on the video transport isembodied in the bursty errors generated by the Markov models.We will examine the impact of lower layer components suchas user mobility, multipath routing, and the MAC layer usingOPNET simulations in Section V.

A. Video Codec Implementations and Parameters

We implemented in software the proposed three video codingschemes on top of the public domain codec [49]. ForRPS, we added our reference picture selection algorithm by ex-tending the RPS option provided by the standard. For LC withARQ, we added a simple rate control algorithm for EL, as ex-plained in Section III-B. For MDMC, the codec was modified toproduce both central and side predictions in the INTER mode,and encodes central and side prediction errors using quantiza-tion parameters and , respectively. More details aboutMDMC can be found in [8].

Error concealment is performed in the decoders when packetsare lost. In the RPS decoder, the conventionalcopy-from-pre-vious-framemethod is used. In the LC decoder, if the BL islost, thecopy-from-previous-framemethod is used. If the ELis lost but the BL is received, the frame is reconstructed usingthe BL only. In the MDMC decoder, the lost information canbe recovered partially from the other description received (seeSection III-C and [8]).

We use the quarter common intermediate format (QCIF),[176 144 Y pixels/frame, 88 72 Cb/Cr pixels/frame]sequence “Foreman” (first 200 frames from the original 30fps sequence) encoded at 10 fps in the performance study ofthe schemes. The encoder generates two substreams with a bitrate of 59 kb/s each. The TMN8.0 [49] rate control methodis used in RPS and LC with ARQ, but the frame layer ratecontrol is disabled. In the simulations using Markov models,RTT is assumed to be three frame intervals. In MDMC,is set to 0.9, and the quantization parameter (, ) isfixed at (8,15) [8], which achieves approximately the same bitrate as the other two schemes. The buffer size of RPS is set to12 frames. If the selected reference frame is not found in thebuffer, the nearest frame is used instead. In all the methods, 5%macroblock level intra-refreshments are used, which has beenfound to be effective in suppressing error propagation for therange of the packet loss rates considered. Each group of blocks(GOB) is packetized into a single packet, to make each packetindependently decodable. In LC with ARQ transmission, theBL is transmitted on the better channel if the two channels areasymmetric. In the following simulations, we only allow a lostBL packet to be retransmitted once.

B. Modeling of Ad Hoc Routes Using Markov Models

In [31], replied routes for a route query are first broken into apool of links and disjoint routes are assembled from the links inthe pool. Motivated by this work, we model a multiple hop wire-less route using the concatenation of a number of links, drawnrandomly from a link pool and each is modeled by a Markovchain [9].

This model has many advantages. The link pool can be easilybuilt using measurement data of ad hoc links. Furthermore, theloss pattern and loss rate are easily controlled. More complexpacket loss processes can also be modeled using semi-Markovmodels. Performance analysis of the video codecs under afullspectrumof loss processes is possible with this technique. Thedisadvantage of this model is that it is a high level model. Detailssuch as bandwidth and end-to-end delay variation, interference,and user mobility cannot be modeled accurately. These issueswill be addressed in the OPNET models in Section IV-C.

C. Simulation Results Using Markov Channel Models

A three-state Markov model was used for each link with thestates representing a “good,” “bad” or “down” status for thelink. The “down” state means the link is totally unavailable. The“good” state has a lower packet loss rate than the “bad” state.The packet loss rates we used are , ,and , for the “down,” the “bad,” and the “good” states,respectively. The transition parameters are chosen to generateloss traces with desired loss rates and mean burst lengths. In


Fig. 4. Average PSNRs of the three schemes with asymmetric paths: Path 1’sloss rate is fixed at 12% and path 2’s loss rate varies from 0.1% to 10%.

Fig. 5. Average PSNRs of the three schemes with asymmetric paths: Path 1’sloss rate is twice of that of path 2.

our simulation, two paths were set up for each connection, andeach path was continuously updated as follows: After every 2 s,two links were chosen randomly from a link pool to constructa new path. For the results reported in Figs. 4–7, the paths usedare disjoint to each other. A video packet can go through a pathcorrectly only when it goes through every link successfully. Foreach case studied in the following, a 400s video sequence is usedby repeatedly concatenating the same video sequence 20 times.The error-free average peak-signal-to-noise ratio (PSNR) of thereceived video frames are 34.14, 33.31, and 33.47 dB for RPS,LC with ARQ, and MDMC, respectively.

The average PSNRs of decoded video sequences under var-ious packet loss rates are given in Figs. 4–7. From the figures, wecan conclude that the best choice depends on the channel char-acteristics, including error rate and error pattern (burst length),the application requirements, including delay constraint, and thedifferences among those channel characteristics (for example,symmetric or asymmetric). Specifically, the following observa-tions can be made from the figures.

1) Effect of Packet Loss Rates:Generally, when the burstlength is not higher than 9 packets, LC with ARQ has the best

Fig. 6. Average PSNRs of the three schemes with symmetric paths: The meanburst length is fixed at 4 packets, while the loss rates varies.

Fig. 7. Average PSNRs of the three schemes with symmetric paths: The lossrates are fixed at 10%, while the mean burst length varies.

performances when the error rate is medium to high. A largeburst of error may make the ARQ scheme less effective (seeFig. 7). RPS has the best performance at very low error rate dueto its coding efficiency.

2) Effect of Path Symmetry:The paths may be symmetricor asymmetric in terms of packet loss rates. Comparing Figs. 5and 6, when the average loss rate on the two paths is equal,LC without ARQ works better in asymmetric channels, whilethe performance of RPS, LC with ARQ, and MDMC is similarwith either asymmetric or symmetric channels. For example, forthe (2%, 4%) point in Fig. 5, LC with ARQ has average PSNRof 32.84 dB, while for the (3%, 3%) point in Fig. 6, LC withARQ has an average PSNR of 32.87 dB. This broadly holdsfor all other points in both figures. With retransmission of lostBL packets on the EL path, the BL path does not need to besignificantly better than the EL path. The fact that the videoquality is insensitive to path symmetry with all three schemes isa blessing: This means that one does not need to provide specialprovisioning in the network to provide at least one high qualitypath.


3) Effect of Error Burst Lengths:From Fig. 7, we observethat in the low to intermediate burst length range considered,the performances of all the three schemes improves graduallyas the burst length increases.3 The reason is when the averagepacket loss rates are the same, a shorter burst length means moreframes have errors, which causes more distortions at the videodecoders; while a larger burst length means less frames have er-rors, which could be remedied with effective error concealment.When the mean error burst length increases, the RPS schemegains most, since its path status prediction method works betterwith longer bursts. Note that this trend is true only up to a cer-tain point. When the burst length increases beyond this point, sothat all packets in two or more consecutive frames are corruptedby an error burst, the PSNR of decoded video will start to dropsharply.

4) Effect of Delay Constraint:When retransmission is al-lowed by the end-to-end delay constraint, LC with ARQ has thebest performance among these three schemes, even only one re-transmission is allowed. We also show in Fig. 4–7 the results forthe LC scheme without retransmissions. This test is used to em-ulate the case when retransmission is not possible, because thedelay constraint of the underlying application does not allow todo so, because it is not feasible to set up a feedback channel, orbecause end-to-end retransmission is not practical (e.g., videomulticast). Without the ARQ protection of BL, the performanceof LC is, as expected, the poorest among all the schemes. Itshows that LC is effective only when the transport layer canprovide efficient error control (e.g., FEC or ARQ) for the BLpackets. Although LC with ARQ has the highest PSNR in thecases we studied, it is the most susceptible to a large end-to-enddelay and imperfect feedback channels.

Note that the performance of the three schemes is also influ-enced by several other factors. For example, the performanceof RPS varies with the RTT as well: the shorter the RTT is, thebetter RPS works. This is because when the RTT is shorter, theRPS encoder will be notified of the corrupted frames earlier andthen it can stop error propagation earlier. Because MDMC doesnot require the set up of a feedback channel, it can be used fora wider range of applications.

V. PERFORMANCESTUDY USING OPNET MODELS

The simplicity of the Markov model enables us to examinethe performance of the proposed schemes over a wide range ofpacket loss patterns. In this section, we use the OPNET modelsto examine the impact of lower layer factors, which are not re-vealed by the Markov model. Using MDMC as an example, weshow how multipath routing, MAC operation and user mobilityaffect video transport.

A. Multipath Routing Using Dynamic Source Routing

DSR is a source routing protocol proposed for mobile ad hocnetworks, where intermediate nodes do not need to maintainup-to-date routing information for forwarding a transit packetsince the packet carries the end-to-end path in its header. It isan on-demand protocol where route discovery is performed fora node only when there is data to be sent to that node [28].

3The burst length in Figs. 4–7 is measured in packets. With the QCIF video,there are nine packets per frame.

Fig. 8. Route updating algorithm.

There are a number of extensions of DSR to multipathrouting. In [22], intermediate nodes are forbidden to reply toroute queries. The destination node receives several copies ofthe same route query (each traverses a possibly different path)within a time window. At the end of the time window, the pathwith the shortest delay and another path which is most disjointwith the shortest path are returned to the source. In [50], inaddition to a shortest path, backup paths from the source nodeand from each intermediate node of the shortest path are foundto reduce the frequency of route request flooding.

We extended DSR to multipath DSR (MDSR) in the fol-lowing way. Each node maintains two routes to a destination.We allow both the destination node and intermediate nodes toreply to a route query. When the destination node replies, it alsocopies the existing routes from its own route cache into the routereply, in addition to the route that the route query traversed. An-other difference with the single path DSR is that when a nodeoverhears a reply which carries a shorter route than the reply itplans to send, it still send its reply to the requesting source. Thisincreases the number of different routes returned, giving thesource a better choice from which to select two maximally dis-joint paths from these replies. The first returned route is used bythe originating packet. Then, the route cache is further updated(or optimized) as new replies for the same query arrive, usingthe algorithm in Fig. 8. This algorithm is a greedy algorithm inthe sense that it always finds the best paths returnedso far. Sincewith DSR, a reply with shorter route usually arrives earlier [28],this heuristic algorithm gives good performance without usinga time window. As compared with [22], the routing delay withMDSR is smaller.

Although MDSR is capable of maintaining more than tworoutes, we only experimented with the two-path version, sincethe results in [50] indicate that the largest improvement isachieved by going from one to two or three paths. The MDSRmodels is built based on the OPNET DSR model [51].


B. OPNET Simulation Setting

Using the OPNET model, we simulate an ad hoc networkwith 16 nodes in a 600 m600 m region. Given the dimen-sions of the region, 16 nodes result in a density that maintainsa connected network for most of the time [52]. Each node israndomly placed in the region initially. We used a version ofthe popularrandom waypointmobility model, where each nodefirst chooses a random destination in the region, then moves to-ward it at aconstantspeed. When it reaches the destination, itpauses for a constant time interval, chooses another destinationrandomly, and then moves toward the new destination [53]. Notethat this is a simplified version of the Random Waypoint model.Since there is no randomness in the nodal speed, the conver-gence problem reported in [54] does not present itself here. Weused a pause time of 1.0 s for all the experiments reported in thispaper. The speed of the nodes varies from 0 m/s to 10 m/s, whichmodels movement of pedestrians or vehicles in city streets.

We use the IEEE 802.11 protocol in the MAC layer workingin the DCF mode. Its physical layer features, e.g., frequencyhopping (FH), are not modeled. The channel has a bandwidthof 1 Mb/s. The transmission range is 250 m. If the sender of apacket is within this range of the receiver, and the sender hassuccessfully accessed the channel during the transmission pe-riod, the packet is regarded as correctly received. The maximumnumber of link layer retransmissions is seven, after which thepacket is dropped. UDP is used in the transport layer. We alsoimplemented part of the RTP functionality [13], such as times-tamping and sequence numbering in our model.

Among the 16 nodes, one is randomly chosen as the videosource and another node is chosen as the video sink, where a5 s playout buffer is used to absorb the jitter in received packets.The video source starts a session using two routes, sending twosubstreams of encoded video (59 kb/s each) to the sink. All othernodes generate background traffic to send to a randomly chosendestination. The interarrival time of the background packets isexponentially distributed with a mean of 0.2 s. The backgroundpacket has a constant length of 512 bits.

C. Simulation Results Using OPNET Models

1) MDSR Performance:First, we examined the per-formance of MDSR. Fig. 9 plots the traces of two routesmaintained by the video source node during a simulation inwhich each node has a speed of 10 m/s. The length of a routeis denoted by the total number of nodes the route traverses,including the source and the destination. Each point in thefigure also means a route update: either a better route is found,or a route in use is broken. It can be seen that the routes areunstable. However, since the nodes are moving around rapidly,new neighbors are found and a new route is discovered shortlyafter the old route is down. This shows that mobility is bothharmful and helpful. The lengths of the routes vary from 0 to 6(5 hops, a little higher than the diameter of the network) duringthe simulation. For most of the simulation period, the routelength is 2, which means direct communication between thesource and the destination, or 3, which means a relay node isused in between. We also plot the number of common nodesbetween two paths. During most of the simulation period, this

Fig. 9. Simulation results of 16 nodes moving in a 600 m� 600 m region at aspeed of 10 m/s. Plotted are the traces of two routes to the video sink maintainedby the video source during the simulation period.

Fig. 10. PSNRs of the received frames with a MDMC codec using two paths.16 nodes move in a 600 m� 600 m region at a speed of 10 m/s. Plotted onthe righty axis are the lost packets per frame. The MSDR algorithm is usedfor route updates. The measured average loss rates of the two substreams are:(3.0%, 3.1%).

number is two, which means the two routes are disjoint exceptfor the common source and destination.

Fig. 9 shows the period when the routes are the longest andmost correlated. At the beginning of the period, the two paths aretwo hops each and are disjoint. Then, they get longer and havemore common nodes, probably indicating that they are movingaway from each other. After the 260th second, they become twohops again after shorter new paths are found. After the 262thsecond, the routes become disjoint again.

2) MPT Versus SPT:Next, we compare the performance ofMPT with single path transport (SPT) in Figs. 10 and 11, wherethe same multistream video coder MDMC is used. For SPT, weused the NIST DSR model [51], which maintains a single pathto a destination, while for MPT we used the MDSR model whichis a multipath routing extension of [51]. We transmit both sub-streams on the same path in the SPT simulations. To alleviate theimpact of bursty errors, we interleave the packets of the two de-scriptions with a two-frame interleaving interval. Description 1


Fig. 11. PSNRs of the received frames with a MDMC codec using a singlepath. 16 nodes move in a 600 m� 600 m region at a speed of 10 m/s. Plotted onthe righty axis are the lost packets per frame. The path is updated using the theNIST DSR model and both substreams are sent on the path using an interleavinginterval of two frames. The measured average loss rates of the two substreamsare: (6.3%,6.4%).

(2) packets for two even (odd) frames is followed by descrip-tion 2 (1) packets for two odd (even) frames. We perform thisexperiment for the 16-node network, where each nodes movesat a speed of 10 m/s. The PSNR traces (using the leftaxis)and loss traces (using the rightaxis) using MDSR and DSRare plotted in Figs. 10 and 11, respectively. It can be seen thatPSNR drops when there is loss in either substream. Also, thedeepest drop occurs when a large burst of loss of one substreamoverlaps with that of the other substream. SPT has higher lossrates than MPT and, therefore, the PSNR curve in Fig. 11 hasmore frequent and severe drops than that in Fig. 10. It is obviousthat SPT has poorer performance than MPT.

This is further illustrated in Fig. 12, which displays thezoomed versions of the loss traces of both simulations. TheSPT traces [Fig. 12(b)] have more frequent packet losses thanthe MPT traces [Fig. 12(a)]. Even worse is that the packetlosses of the substreams are strongly correlated in the SPT case.This has the most negative impact on the MDMC performance.Although we interleaved the substreams before transmitting,the burst length is too long, rendering the interleaving ineffec-tive. A larger interleaving interval may help but at the cost of alarger end-to-end delay.

3) Impact of Mobility: We also examined the impact ofmobility on video transport using MDMC as an example.Fig. 13 shows the PSNRs of the received video frames whenthe nodes are stationary. The PSNR curve in Fig. 13 is verystable, with only a few narrow drops. From the error tracebelow, we can see that the losses are mostly random, i.e., theerror burst length is one packet for most of the time. Whennodes begin to move around at a speed of 10 m/s, the PSNRcurve in Fig. 10 is much worse with many more drops. Thelargest drops occur at the 500th, 2000th, and 2500th frames.4

We conjecture that during these periods the source node anddestination node were either far away from each other, or were

4Note that the frame rate is 10 frames/s. The frame number corresponding tothe 250th second in Fig. 9 is 2500 in Fig. 10.

(a)

(b)

Fig. 12. Zoomed lost packet per frame traces of two substreams. (a) MPT case(corresponding to Fig. 10). (b) SPT case (corresponding to Fig. 11).

Fig. 13. PSNRs of the received frames with a MDMC codec. 16 nodes in a600 m� 600 m region. Plotted on the righty axis are the lost packets per frame.The nodes are stationary.

in a hot-spot. We can see that in both figures, the valleys of thePSNR curve match the loss bursts drawn below. In Fig. 10, theloss bursts in this plot match the two routes’ longest and mostcorrelated periods in Fig. 9. Mobility clearly has a negativeeffect on video transport.

Fig. 14 shows the mean packet loss rates and the mean packetloss burst lengths of the two substreams during the simulationswhen the mobile speed varies from 0 m/s to 10 m/s. Fig. 15 is theresulting average PSNR for different speeds using the MDMCcodec. There are several interesting observations.

1) The two routes maintained by MDSR are relatively sym-metric in their error characteristics. Recall from Fig. 8


(a)

(b)

Fig. 14. Loss characteristics versus mobile speed for both MPT and SPTOPNET simlations. (a) Average packet loss rate. (b) Average error burst length.

Fig. 15. Average PSNRversus node speed for the MDMC scheme from theOPNET simulations.

that the algorithm does not order the two paths MDSRmaintains. This is suitable for MDMC since the two de-scriptions are equivalent in importance on video quality.If LC with ARQ is used, an algorithm that always putsthe shortest path in cache 0 (which is used by the BL) ispreferred.

2) MDSR effectively reduces both the mean packet loss rateand the mean packet loss burst length. When the speedis 10 m/s, the average loss rate for MDSR is about 3%,while that of DSR is about 6.4%. The mean burst lengthof MDSR at 10 m/s is about 11, while that of DSR isabout 20.

3) When the nodes are stationary, the mean burst length ofMDSR is 1. Packet loss in this case is mainly caused byfailure in accessing the channel. When nodes are mobile,the links are more frequently broken because of nodalmobility, and the loss characteristics change from randomloss to bursty loss. We conjecture that the burst lengthdepends on the time scale of the routing protocol and therate of change in the topology.

4) Somewhat counterintuitive is that as speed increases,the average loss rate becomes stable. Furthermore, thehighest mean burst length occurs at 4 m/s instead of10 m/s. The mean PSNR in Fig. 15 shows the same trend.When speed increases, the mean PSNR first drops, thenbecomes stable. Note that at 4 m/s, the mean burst lengthis about 40 packets, which correspond to more than fourframes. At this high burst length (and the correspondinghigh loss rate), the general trend that the PSNR increaseswith the burst length is not true any more. In fact, thelikelihood that both paths experience packet lossessimultaneously is quite high, which leads to a significantdrop in the decoded video quality.

These observations further verify our previous observation thatmobility is both harmful and helpful. During the initial increasein mobility, routes break down more easily, which leads to anincrease in the mean packet loss rate and a drop in the PSNR.As speed further increases, new topologies are more quicklyformed and new routes are more quickly established. A hot spotin the region, where nodes cluster and compete for the channel,is more quickly dispersed. As speed increases, the period of timea node remains disconnected is smaller. The turning point (4 m/sin Figs. 14 and 15 ) is determined by the node density in theregion and the transmission range. We conjecture that similarphenomenon exists for other scenarios with a different numberof nodes or a different transmission range, given that the nodedensity is high enough to maintain a connected network for mostof the time. When the nodal speed increases even further, therouting process would be unable to track the quickly changingtopology. Therefore, drops in the average PSNR is expected.

VI. A N AD HOC MULTIPATH VIDEO STREAMING TESTBED

To validate the feasibility of video transport over ad hocnetworks and evaluate the achievable video quality withtoday’s technology, we implemented an ad hoc multipathvideo streaming testbed. The implementation and experimentalresults are reported in the following.

A. Setup of the Testbed

The testbed is implemented using four IBM Thinkpad note-books with 802.11b cards. Fig. 16 shows the network view of thetestbed. The notebooks were placed at (or moved around) variouslocations in the Library/CATT building at Polytechnic Uni-versity. IBM High Rate Wireless LAN cards are used workingin the DCF mode with a channel bandwidth of 11 Mb/s. Thecorridors are about 30 m 60 m. In the building, there isinterference from IEEE 802.11 access points (AP) and otherelectronic devices (e.g., microwave ovens). Nodesand are,respectively, the video source and sink, while nodesand


(a) (b)

Fig. 16. Experiment scenarios for the testbed. (a) Line-of-sight. (b) Behind thewalls.

are the relays. Since there are only four nodes in this net-work, we use static routing to force the use of two-hop routes.Dynamic routing will be implemented in a future version.

The system is built on Microsoft Windows 2000. We im-plemented the timestamping, sequence numbering, and QoSfeedback functions in the application layer. For LC with ARQ,limited retransmission for BL is implemented in the applicationlayer as well. UDP sockets are used at the transport layer. Atraffic allocator dynamically allocates packets to the two paths.The video sink maintains a playout buffer, using a hash tabledata structure. Typically video streaming applications use aplayout buffer of a few seconds to smooth the jitter in incomingpackets. We chose a playout buffer of 2 s for this network. Tosupport interactive applications, we also experimented with a300 ms playout delay.

The implementations of LC with ARQ and MDMC codec dis-cussed in Section IV-A are used. We did not use the RPS codecsince it does not support streaming of pre-encoded video. Forboth schemes used, the testbed performs off-line encoding, butthe received video frames are decoded in real time and displayedon the screen of node .

B. Experimental Results

We examined the performance of the system in the scenariosshown in Fig. 16. The average PSNRs of the received framesusing the two schemes are presented in Table II and Table III,respectively. For comparison with the Markov simulationstudies presented in Section IV, the last row of each table liststhe average PSNRs obtained from Figs. 4–7. Each testbedvalue in the table is the average over an experiment lasting for10–15 min. In all the tests using LC with ARQ, path 1 is usedfor the BL.

The results in Table II are consistent with the Markov simu-lation results presented in Section IV-C. The minor differencebetween the last two rows of Table II are caused by the differ-ences in the actual loss rates of the testbed experiments and thecorresponding Markov simulations, and the differences in theloss patterns (the experimental loss processes may not be Mar-kovian). An interesting observation is that for the test scenarioin Fig. 16(a), when the loss rates for both substreams are verylow, MDMC has a higher average PSNR (33.11 dB) than LCwith ARQ (32.24 dB). This is also shown in Figs. 5 and 6.

TABLE IIAVERAGE PSNRS OF DECODED FRAMES: MDMC TESTBED

EXPERIMENTS/MARKOV SIMULATIONS

TABLE IIIAVERAGE PSNRS OF DECODED FRAMES: LC WITH ARQ TESTBED

EXPERIMENTS/MARKOV SIMULATIONS

The results in Table III clearly show that ARQ effectivelyreduces the BL packet loss rate in all the experiments. Thesereductions account for improved video quality. For example,all lost BL packets in the test of Fig. 16(a) were successfullyrecovered, resulting in a BL loss rate of 0%. The average PSNRfor this test is the highest among all the LC with ARQ experi-ments. However, this is not true for the mean burst length. Sinceeach BL packet has its deadline imposed by the playout delay,ARQ is more effective in recovering short error bursts. Duringan experiment, many short error bursts are recovered, reducingthe number of error bursts. But for an error burst comparableor greater than the playout delay, only a small portion of itis successfully retransmitted. Therefore, sometimes the meanburst length of BL increases when ARQ is used. Recall inFig. 7, we showed that given the same average loss rate, PSNRimproves when the mean burst length increases. The increasedmean burst length, combined with reduced average BL lossrate, contributes to the increased video quality. During theexperiments with LC with ARQ, we observed short periodsof badly corrupted frames, followed by a long period of highquality frames. The ARQ success ratio in the table is definedto be the ratio of the number of successfully retransmittedBL packets and the number of all lost BL packets. This ratiodecreases as the loss rates of both paths increase and as theplayout delay decreases.

For the experiments reported here, the packet delay anddelay jitter are both very low, because there is low backgroundtraffic and the bit rates of the video substreams are also low. Forthis reason, very few packets are dropped because of latenesseven with a 300 ms playout delay, and the video equalityobtained with a 300 ms playout delay is similar to that with a2 s playout delay, with both schemes. When the system load


Fig. 17. A screenshot of the testbed GUI during a MDMC experiment.

is higher, the 300 ms playout delay is likely to yield worseperformance.

In the experiments, LC with ARQ performs better thanMDMC, except for the very low loss cases, which is consistentwith the Markov model simulation results in Figs. 4–7. How-ever, as compared with the results in Figs. 4–7, LC with ARQyields lower average PSNR under similar loss rates. In fact, asuccessful retransmission requires: 1) successful and timelydelivery of the NACK and 2) successful delivery of the retrans-mitted BL packet before its deadline. Recall that in the Markovmodel simulations, we assume a perfect feedback channeland the paths were chosen to be disjoint. These assumptions

ensure a high ARQ success ratio. Additionally, in the testbed,NACKs are sent on the same two paths as the video packets.The end-to-end delay is not fixed as in the Markov modelsimulations. These further reduce the degree of path diversityand the effectiveness of the ARQ algorithm. For MDMC, nofeedback is necessary. As a result, its performance is moreconsistent with the Markov simulation results in Figs. 4–7.

Fig. 17 is a screenshot of the MDMC testbed during an exper-iment. The upper left part of the GUI displays the received videoand the network view of the testbed. The transport related sta-tistics (loss rates, jitter, receiver buffer occupancy, etc.) and thevideo codec related attributes (frame rate, format, bit rates, etc.)


are displayed at the upper right part of Fig. 17. The two windowsin the center display the packet loss traces of the two substreamsfor each frame. The lower part of Fig. 17 is the PSNR trace ofthe received video, which illustrates how video quality is im-paired by packet losses of both substreams and how the MDMCdecoder recovers from the packet losses.

VII. CONCLUSION

Enabling video transport over ad hoc networks is morechallenging than over other wireless networks, both becausead hoc paths are highly unstable and compressed video issusceptible to transmission errors. However, multiple paths inan ad hoc network can be exploited as an effective means tocombat transmission errors. Motivated by this observation, wechose three representative video coding schemes, all based onMCP used in modern video coding standards, and show how toadapt these schemes with MPT to enable video transport overad hoc networks.

In this paper, we highlight the close interaction between thevideo codec and the transport layer. Results presented suggestthat if a feedback channel can be set up, the standard H.263coder with its RPS option can work quite well. Additionally,if delay caused by one retransmission is acceptable, layeredcoding with the BL protected by ARQ is more suitable. MDCis the best choice when feedback channels cannot be set up, orwhen the loss rates on the paths are not too high. A comparisonof the schemes from the perspective of video coding, e.g., framememory required at the coder and decoder, source coding redun-dancy, and on-line/off-line coding is presented as well. UsingOPNET models, we extended DSR to support multipath routing.The impact of multipath routing and mobility is investigated.

To further verify the feasibility of video transport over adhoc networks, we implemented an ad hoc multipath videostreaming testbed using four notebook computers. Our testsshow that acceptable quality streaming video is achievablewith both LC with ARQ and MDMC, in the range of video bitrate, background traffic, and motion speed examined. Togetherwith simulation results using the Markov path model and theOPNET models, our studies demonstrate the viability of videotransport over ad hoc networks using multipath transport andmultistream coding.

We should note that further improvements could be made foreach component in the proposed framework. For example: 1) thevideo codec parameters could be further tuned and optimized inthe rate distortion sense, given the path conditions; 2) packetsfrom all the substreams could be dispersed to the multiple pathswith a more sophisticated algorithm to maximize the benefit ofMPT; and 3) the ad hoc networks simulated in Markov modeland OPNET models, and the testbed are relative small. It wouldbe very interesting to see how the performance scales for a largerad hoc network. These are still open research problems that areworth investigating in future work.

ACKNOWLEDGMENT

The authors would like to thank the anonymous reviewersfor their helpful comments which improved the quality of this

paper. We also thank the Wireless Communications Technolo-gies Group of the National Institute of Standards and Tech-nology for making the OPNET DSR model available in thepublic domain.

REFERENCES

[1] Draft Standard for Wireless LAN: Medium Access Control (MAC) andPhysical Layer (PHY) Specifications IEEE, IEEE Standard 802.11, July1996.

[2] C. R. Lin and M. Gerla, “Asynchronous multimedia multihop wirelessnetworks,” inProc. IEEE/ACM INFOCOM, Kobe, Japan, Apr. 1997, pp.118–125.

[3] Y. Wang, S. Wenger, J. Wen, and A. K. Katsaggelos, “Error resilientvideo coding techniques,”IEEE Signal Processing Mag., vol. 17, pp.61–82, July 2000.

[4] Y. Wang and Q.-F. Zhu, “Error control and concealment for video com-munication: A review,”Proc. IEEE, vol. 86, no. 5, pp. 974–997, May1998.

[5] N. Gogate, D. Chung, S. S. Panwar, and Y. Wang, “Supportingimage/video applications in a multihop radio environment using routediversity and multiple description coding,”IEEE Trans. Circuit Syst.Video Technol., vol. 12, no. 9, pp. 777–792, Sept. 2002.

[6] S. Lin, S. Mao, Y. Wang, and S. S. Panwar, “A reference pictureselection scheme for video transmission over ad hoc networks usingmultiple paths,” in Proc. IEEE ICME, Tokyo, Japan, Aug. 2001,pp. 96–99.

[7] S. Mao, S. Lin, S. S. Panwar, and Y. Wang, “Reliable transmission ofvideo over ad hoc networks using automatic repeat request and multipathtransport,” inProc. IEEE Fall VTC, Atlantic City, NJ, Oct. 2001, pp.615–619.

[8] Y. Wang and S. Lin, “Error resilient video coding using multiple descrip-tion motion compensation,”IEEE Trans. Circuits Syst. Video Technol,vol. 12, no. 6, pp. 438–452, 2002.

[9] E. N. Gilbert, “Capacity of a bursty-noise channel,”Bell Syst. Tech. J.,vol. 39, no. 9, pp. 1253–1265, Sept. 1960.

[10] OPNET Modeler. OPNET Tech., Inc. [Online]. Available: http://www.mil3.com

[11] S. Wenger, “Video redundancy coding inH:263+,” in WorkshopAudio-Visual Services for Packet Networks, Aberdeen, Scotland,Sept. 1997.

[12] M. Khansari, A. Jalali, E. Dubois, and P. Mermelstein, “Low bit-ratevideo transmission over fading channels for wireless microcellular sys-tems,” IEEE Trans. Circuit Syst. Video Technol., vol. 6, pp. 1–11, Feb.1996.

[13] H. Schulzrinne, S. Casner, R. Frederick, and V. Jacobson. RTP: A trans-port protocol for realtime applications. presented at IETF request forcomments 1889. [Online]. Available: http://www.ietf.org

[14] V. K. Goyal, “Multiple description coding: Compression meets the net-work,” IEEE Signal Processing Mag., vol. 18, pp. 74–93, Sept. 2001.

[15] A. Reibman, H. Jafarkhani, Y. Wang, M. Orchard, and R. Puri, “Multipledescription coding for video using motin compensated prediction,” inProc. IEEE ICIP, Kobe, Japan, Oct. 1999, pp. 837–841.

[16] J. G. Apostolopoulos, “Error-resilient video compression through theuse of multiple states,” inProc. IEEE ICIP, Vancouver, BC, Canada,Sept. 2000, pp. 352–355.

[17] M. de Cordeiroet al., “Establishing a trade-off between unicast and mul-ticast retransmission modes for reliable multicast protocols,” inProc.Int. Symp. Modeling, Analysis, Simulation of Computer Telecommuni-cation Systems, 2000, pp. 85–91.

[18] N. F. Maxemchuck, “Diversity routing,” inProc. IEEE ICC, vol. 1, SanFrancisco, CA, June 1975, pp. 10–41.

[19] T. Wu and R. Lau, “A class of self-healing ring architectures for SONETnetwork applications,”IEEE Trans. Commun., vol. 40, pp. 1746–1756,Nov. 1992.

[20] R. R. Stewart and Q. Xie,Stream Control Transmission Protocol: A Ref-erence Guide. Reading, MA: Addison-Wesley, 2001.

[21] D. Sidhu, R. Nair, and S. Abdallah, “Finding disjoint paths in networks,”in Proc. ACM SIGCOMM, Zurich, Switzerland, Sept. 1991, pp. 43–51.


[22] S.-J. Lee and M. Gerla, “Split multipath routing with maximally disjointpaths in ad hoc networks,” inProc. IEEE ICC, Helsinki, Finland, June2001, pp. 3201–3205.

[23] P. Papadimitratos, Z. J. Haas, and E. G. Sirer, “Path set selection in mo-bile ad hoc networks,” inProc. ACM MOBIHOC, Lausanne, Switzer-land, June 2002, pp. 1–11.

[24] A. Nasipuri, R. Castañeda, and S. R. Das, “Performance of multipathrouting for on-demond protocols in mobile ad hoc networks,”MobileNetworks and Applications, vol. 6, no. 4, pp. 339–349, 2001.

[25] S. Nelakuditi, Z.-L. Zhang, and R. P. Tsang, “Adaptive proportionalrouting: A localized QoS routing approach,” inProc. IEEE INFOCOM,Tel-Aviv, Israel, Mar. 2000, pp. 1566–1575.

[26] J. Moy. OSPF version 2. presented at IETF Request for Comments 2328.[Online]. Available: http://www.ietf.org

[27] E. Gustafsson and G. Karlsson, “A literature survey on traffic disper-sion,” IEEE Network, pp. 28–36, Mar./Apr. 1997.

[28] D. B. Johnson, D. A. Maltz, Y.-C. Hu, and J. G. Jetcheva, “The dynamicsource routing protocol for mobile Ad Hoc networks,” in IETF InternetDraft (draft-ietf-manet-dsr-03.txt), Oct. 1999.

[29] C. E. Perkins, E. M. Royer, and S. R. Dos, Ad hoc on-Demand Dis-tance Vector (AODV) Routing, IETF Internet Draft (draft-ietf-manet-aodv-06.txt), July 2000.

[30] Z. J. Haas and M. R. Pearlman, “The zone routing protocol (ZRP) for adhoc networks,” in IETF Internet Draft (draft-ietf-manet-zone-zrp02.txt),June 1999.

[31] M. Pearlman, Z. J. Hass, P. Sholander, and S. S. Tabrizi, “On the impactof alternate path routing for load balancing in mobile ad hoc networks,”in Proc. MobiHOC, Aug. 2000, pp. 3–10.

[32] N. Gogate and S. S. Panwar, “On a resequencing model for high speednetworks,” inProc. IEEE INFOCOM, June 1994, pp. 40–47.

[33] Y. Nebat and M. Sidi, “Resequencing considerations in parallel down-loads,” inProc. IEEE INFOCOM, June 2002, pp. 1326–1336.

[34] N. Gogate and S. S. Panwar, “Supporting applications in a mobile mul-tihop radio enviroment using route diversity: I. Nonrealtime data,” inProc. IEEE ICC, June 1998, pp. 802–806.

[35] D. S. Phatak and T. Goff, “A novel mechanism for data streaming acrossmultiple IP links for improving throughput and reliabilit in mobile envi-ronments,” inProc. IEEE INFOCOM, June 2002, pp. 773–782.

[36] H.-Y. Hsieh and R. Sivakumar, “A transport layer approach for achievingaggregate bandwidths on multi-homed mobile hosts,” inProc. ACM Int.Conf. Mobile Computer Networking, Sept. 2002, pp. 83–95.

[37] N. Gogate and S. S. Panwar, “Supporting video/image applications in amobile multihop radio environment using route diversity,” inProc. IEEEICC, June 1999, pp. 1701–1706.

[38] J. G. Apostolopoulos, “Reliable video communication over lossy packetnetworks using multiple state encoding and path diversity,” inProc.SPIE Conf. Visual Communication Image Processing, Jan. 2001, pp.392–409.

[39] J. G. Apostolopoulos, T. Wong, W. Tan, and S. Wee, “On multiple de-scription streaming in content delivery networks,” inProc. IEEE IN-FOCOM, June 2002, pp. 1736–1745.

[40] T. Nguyen and A. Zakhor, “Path diversity with forward error correction(PDF) system for packet switched networks,” in IEEE INFOCOM, SanFrancisco, CA, Apr. 2003.

[41] Y. J. Liang, E. G. Steinbach, and B. Girod, “Multi-stream voice overIP using packet path diversity,” inProc. IEEE Multimedia Signal Pro-cessing Workshop, Sept. 2001, pp. 555–560.

[42] Y. J. Liang, E. Setton, and B. Girod, “Channel-adaptive video streamingusing packet path diversity and rate-distortion optimized reference pic-ture selection,” in The IEEE Multimedia Signal Processing Workshop,St. Thomas, U.S. Virgin Islands, Dec. 2002, pp. 420–423.

[43] V. Jacobson, “Congestion avoidance and control,”Comput. Commun.Rev., vol. 19, no. 4, pp. 314–329, Aug. 1988.

[44] D. A. Maltz, J. Broch, and D. B. Johnson, “Lessons from a full-scalemulti-hop wireless ad hoc network testbed,”IEEE Pers. Commun., vol.8, pp. 8–15, Feb. 2001.

[45] H. Lundgrenet al., “A large-scale testbed for reproducible ad hoc pro-tocol evaluations,” inProc. IEEE WCNC, 2002, pp. 412–418.

[46] W. Kellerer, E. Steinbach, P. Eisert, and B. Girod, “A real-time internetstreaming media testbed,” inProc. IEEE ICME, Aug. 2002, pp.453–456.

[47] W. R. Stevens,TCP/IP Illustrated, Volume I: The Protocols. Reading,MA: Addison-Wesley, 1994.

[48] “Video Coding for Low Bit Rate Communication,” ITU-T Recommen-dation H.263, 1998.

[49] H:263+ Codec. University of British Columbia. [Online]. Available:ftp://dspftp.ece.ubc.ca

[50] A. Nasipuri and S. R. Das, “On-demand multipath routing for mobileAd hoc networks,” inProc. ICCCN, 1999, pp. 64–70.

[51] OPNET DSR Model. The Wireless Communications TechnologiesGroup of the National Institute of Standards and Technology. [Online].Available: http://w3.antd.nist.gov/wctg/prd_dsrfiles.html

[52] T. K. Philips, S. S. Panwar, and A. N. Tantawi, “Connectivity propertiesof a packet radio network model,”IEEE Trans. Inform. Theory, vol. 35,pp. 1044–1047, Sept. 1989.

[53] J. Broch, D. A. Maltz, D. B. Johnson, Y.-C. Hu, and J. Jetcheva, “Aperformance comparison of multi-hop wireless ad hoc network routingprotocols,” inProc. ACM/IEEE Int. Conf. Mobile Computers and Net-working, 1998, pp. 85–97.

[54] J. Yoon, M. Liu, and B. Noble, “Random waypoint considered harmful,”in Proc. INFOCOM, San Francisco, CA, Apr. 2003, pp. 1312–1321.

Shiwen Mao(S’99) received the B.S. degree in elec-trical engineering, the B.E. degree in enterprise man-agement, the M.S. degree in electrical engineeringfrom Tsinghua University, Beijing, China, in 1994and 1997, respectively, and the M.S. degree in systemengineering from Polytechnic University, Brooklyn,NY, in 2000, where he is currently working towardthe Ph.D. degree.

From 1994 to 1995, he was a Research Assistantat the State Key Lab on Microwave and Digital Com-munications, Beijing, China. From 1997 to 1998, he

was a Research Member of IBM China Research Laboratory. In the summerof 2001, he was a Research Intern in Avaya Labs-Research, Holmdel, NJ. Hisresearch interests include real-time transport in the Internet and wireless net-works, performance of wireless ad hoc networks, and queueing analysis.

Shunan Lin (S’00) received the B.S. degree fromUniversity of Science and Technology of China,Hefei, and the M.S. degree from the Institute ofAutomation, Chinese Academy of Sciences, bothin electrical engineering. He is working toward thePh.D. degree at Polytechnic University, Brooklyn,NY.

From 2000 and 2002, he was an InternshipStudent in Microsoft Research, Beijing, China, andMitsubishi Electric Research Lab, Murray Hill, NJ,respectively. His research interests are in the field

of video singal processing, transmission, and motion estimation.

Shivendra S. Panwar (S’82–M’85–SM’00) re-ceived the B.Tech. degree in electrical engineeringfrom the Indian Institute of Technology, Kanpur, in1981, and the M.S. and Ph.D. degrees in electricaland computer engineering from the Universityof Massachusetts, Amherst, in 1983 and 1986,respectively.

He is a Professor in the Electrical and ComputerEngineering Department, Polytechnic University,Brooklyn, NY. He is currently the Director of theNew York State Center for Advanced Technology

in Telecommunications (CATT). He spent the summer of 1987 as a VisitingScientist at the IBM T. J. Watson Research Center, Yorktown Heights, NY, andhas been a Consultant to AT&T Bell Laboratories, Holmdel, NJ. His researchinterests include the performance analysis and design of networks. Currentwork includes protocol analysis, traffic and call admission control, switchperformance, and multimedia transport over wireless networks.

Dr. Panwar has served as the Secretary of the Technical Affairs Council ofthe IEEE Communications Society (1992–1993) and is a Member of the Tech-nical Committee on Computer Communications. He is a coeditor ofNetworkManagement and Control, Vol. IIandMultimedia Communications and VideoCoding(New York: Plenum, 1994 and 1996).


Yao Wang (M’90–SM’98) received the B.S. andM.S. degrees in electronic engineering from Ts-inghua University, Beijing, China, in 1983 and 1985,respectively, and the Ph.D. degree in electrical andcomputer engineering from University of Californiaat Santa Barbara, in 1990.

Since 1990, she has been with the faculty of Poly-technic University, Brooklyn, NY, and is presentlyProfessor of electrical and computer engineering.She was on sabbatical leave at Princeton University,Princeton, NJ, in 1998, and was a Visiting Professor

at University of Erlangen, Erlangen, Germany, in the summer of 1998. Shewas a Consultant with AT&T Labs—Research, Middletown, NJ, formerlyAT&T Bell Laboratories, from 1992 to 2000. Her research areas include videocommunications, multimedia signal processing, and medical imaging. She isthe leading author of the textbookVideo Processing and Communicationsandhas published over 100 papers in journals and conference proceedings.

Dr. Wang has served as an Associate Editor for IEEE TRANSACTIONS ON

MULTIMEDIA and IEEE TRANSACTIONS ONCIRCUITS AND SYSTEMS FORVIDEO

TECHNOLOGY. She received New York City Mayor’s Award for Excellence inScience and Technology in the Young Investigator Category in 2000.

Emre Celebi (S’96) received the B.Sc. degreein computer engineering and the M.Sc. degreein system and control engineering from BogaziciUniversity, Istanbul, Turkey, in 1998 and 2001,respectively. He is currently a Research Fellow andis working toward the Ph.D. degree in computerscience at Polytechnic University, Brooklyn, NY.

His research interests include wireless networks,performance analysis, evolutionary and adaptiveprotocols.

Video transport over ad hoc networks: multistream coding ...catt.poly.edu/~panwar/publications/video...

Documents

Transcript of Video transport over ad hoc networks: multistream coding ...catt.poly.edu/~panwar/publications/video...